Vector-wise Quantization: Methods & Insights

• Vector-wise Quantization is a discretization method that processes entire high-dimensional vectors to capture intra-vector dependencies and achieve efficient data compression.
• It underpins practical applications like deep neural network compression and generative modeling by enabling higher compression rates at fixed distortion levels.
• Mathematical frameworks such as k-means, convex optimization, and specialized strategies like sign-splitting and polar parameterization are crucial for its effective implementation.

Vector-wise quantization refers to discretization strategies that operate on entire high-dimensional vectors (subvectors or full weight/activation vectors), as opposed to isolated scalar components. Unlike scalar quantization, which treats coordinates independently, vector-wise methods model intra-vector structure, enable higher compression rates at fixed distortion, and underpin a variety of state-of-the-art approaches in deep generative modeling and neural network compression.

1. Mathematical Foundations and Core Algorithms

Vector-wise quantization (VQ) partitions a collection of $N$ input vectors $\mathbf{x}_i\in\mathbb{R}^d$ into $M$ clusters or subsets $\{\mathcal S_m\}_{m=1}^M$, associating each with a representative codeword (or class-distribution in supervised variants). Classical VQ, as formalized in $k$-means, seeks minimization of intra-cluster distortion: $$\min_{C, a}\;\sum_{i=1}^N \|\mathbf{x}_i - c_{a(i)}\|_2^2$$ where $C = \{c_1, ..., c_K\}$ is the codebook, and $a(i)$ maps each input to a codeword (Li et al., 11 Mar 2025).

Supervised extensions, such as KL-minimizing VQ, minimize a labeling-information-preserving loss: $$\min_{\mathcal S_1,\dots,\mathcal S_M} \sum_{i=1}^N D_{\mathrm{KL}}\left(p(\cdot|\mathbf{x}_i)\,\|\,p(\cdot|\mathcal S_{\mu(i)})\right)$$ where $p(y|\mathbf{x}_i)$ encodes class-label statistics (Yang et al., 2015).

VQ can also be reframed in convex optimization terms, as in Soft Convex Quantization (SCQ): $$\alpha^*(x) = \arg\min_{\alpha\in \Delta_K} \|x-C\alpha\|_2^2$$ with $x$ approximated as a convex combination over codewords: this yields fully differentiable quantization and combats codebook collapse (Gautam et al., 2023).

2. Specialized VQ Strategies and Theoretical Advances

2.1. Sign-Splitting, Polar, and Calibration-Free VQ

Limiting cluster updates to codebook indices restricts fine-tuning. Sign-Splitting VQ (SSVQ) decouples sign bits and magnitudes, introducing latent, learnable sign variables, with a progressive freezing schedule to ensure convergence, effectively mitigating the "gradient dominance" induced by shared codeword updates (Li et al., 11 Mar 2025).

Polar Coordinate Decoupled VQ (PCDVQ) recognizes the disproportionate impact of directional quantization error versus magnitude error, especially in ultra-low-bit scenarios (e.g., 2 bit/8-dim). By separating polar parameterizations, PCDVQ performs independent codebook assignment and optimization for direction and magnitude, allocating more bits to angular components, and aligns codebook construction to the empirical direction (sphere-packing) and magnitude (Gaussian chi-distribution) statistics of neural weights (Yue et al., 5 Jun 2025).

NSNQuant leverages a double normalization (normalize–shift–normalize) and Hadamard transform to standardize distributions, enabling generically applicable global codebooks and eliminating calibration data dependence, with robust 1–2 bit quantization for LLM KV caches (Son et al., 23 May 2025).

2.2. Feature- and Task-Aware VQ Objective

Some VQ methods directly optimize the output distributional match (e.g., class label histograms or layer activations). For instance, KL-VQ employs an EM-style optimization to synthesize codebook clusters that minimize the aggregate KL divergence between empirical per-point class conditionals and aggregate cluster label distributions (Yang et al., 2015). In high-variance recognition tasks, such preservation of class-conditional information yields significant improvements (4–5%) over unsupervised quantizers.

Activation-aware methods instead minimize the output reconstruction error on in-distribution network activations, rather than on weights, leading to improved post-training quantization results and fewer accuracy trade-off penalties (Stock et al., 2019).

3. Stochastic, Differentiable, and Information-Theoretic Approaches

Not all VQ variants constrain each input to a single codeword. SCQ generalizes the assignment to the entire probability simplex, effectively solving a small convex program for each vector, and supporting gradients via implicit differentiation of the KKT system. Empirically, this increases codebook utilization and yields dramatic reductions (10–100×) in quantization error compared to hard VQ (Gautam et al., 2023).

DiVeQ and its "space-filling" extension (SF-DiVeQ) reinterpret quantization as the addition of a parameterized random distortion vector, allowing gradients (via the reparameterization trick) to flow through hard assignments, and even assigning to points along codeword-segment curves to ensure full codebook usage (Vali et al., 30 Sep 2025).

Convex-hull and lattice-based approaches (e.g., vqSGD, dual quantization) provide unbiased estimators and uniform error guarantees, with precise information-theoretic rate–distortion characterizations. The vqSGD construct, for instance, utilizes convex combination sampling to ensure unbiasedness, with provable lower and upper bounds on communication, achieving optimal rates up to a constant factor in distributed gradient quantization (Gandikota et al., 2019). Dual quantization (Delaunay-based) guarantees intrinsic second-order stationarity, supporting quadrature error bounds without requiring strictly optimal grids (Pagès et al., 2010).

4. Application Domains and Empirical Impact

Vector-wise quantization is central to nearly all practical large-scale neural network compression and discrete generative modeling:

• Deep Model Compression: VQ outperforms scalar/element-wise quantization in low-bit (2–3 bits) regimes, preserving top-1 accuracy within 1–2 points even for billion-parameter models (ResNet, LLaMA, ViT, DiT) (Li et al., 11 Mar 2025, Yue et al., 5 Jun 2025, Son et al., 23 May 2025, Stock et al., 2019, Egiazarian et al., 2024).
• Diffusion Models and Generative Architectures: VQ enables compressing diffusion U-Nets and DiTs from >2B parameters at 4.15→3.15 bits/weight with negligible FID, CLIP, human preference differences relative to full precision (Egiazarian et al., 2024, Deng et al., 2024). Block-wise and activation-aware calibration further improves robustness.
• LLM KV Caches: Calibration-free VQ methods, notably NSNQuant and TurboQuant, quantize KV caches to ≤2 bits per value in LLMs, preserving perplexity and throughput across domain shifts (Son et al., 23 May 2025, Zandieh et al., 28 Apr 2025).
• Distributed Optimization: VQ-based gradient schemes in vqSGD guarantee unbiased estimation, low communication, and differential privacy simultaneously (Gandikota et al., 2019).
• Reinforcement Learning & Communication Bottlenecks: Dynamic VQ architectures (DVQ), with per-input codebook selection, adapt quantization tightness as required by context complexity in multi-agent or visually-rich environments (Liu et al., 2022).

5. Optimization Techniques and Workflow

A canonical optimization pipeline for vector-wise quantization involves:

1. Partitioning Weights or Latents: Reshape network parameters/activations into groups or blocks of dimension $d$ (e.g., $d=4, 8$).
2. Codebook Construction: Learn codebooks $C$ via (weighted) $k$-means, additive/multi-codebook optimization, or analytic/spherical design (e.g., PCDVQ for direction).
3. Assignment Optimization: Assign each vector/subvector to its nearest codeword(s), potentially via beam search, convex assignment (SCQ), or with soft candidate sets (VQ4DiT).
4. Specialized Regularization: Apply explicit regularizers for codebook usage (prior KL), stochastic assignment masking, or progressive sign-freezing (SSVQ), as suited (Zhang et al., 2023, Li et al., 11 Mar 2025).
5. Fine-tuning and Calibration: Optionally perform task-distillation or zero-data feature-matching to further align codebooks and assignments to layer-specific distributions or block outputs (Deng et al., 2024, Egiazarian et al., 2024).

6. Quantitative Results and Empirical Trade-offs

Vector-wise quantization consistently outperforms scalar quantization in preservation of network accuracy and generative fidelity at aggressive bit-rates:

Method	Bits/Weight	Top-1 Accuracy / Metric	Reference
SSVQ (DeiT-tiny)	21× comp.	25% (vs. 13% for conventional VQ)	(Li et al., 11 Mar 2025)
PCDVQ (LLaMA-2-7B, 2 bit)	2	QA avg 58.6% (vs 58.13% VPTQ)	(Yue et al., 5 Jun 2025)
NSNQuant (KV caches)	2	Perplexity 9.08 (vs. 9.75 for CQ)	(Son et al., 23 May 2025)
VQDM (SDXL, 3.15 bits)	3.15	FID 19.18 (vs 19.78 for 4-bit PTQ4DM)	(Egiazarian et al., 2024)

Notably, these methods maintain or exceed the practical accuracy or generation metrics achieved by traditional scalar quantizers, especially as bits per vector decrease below 4, due to superior modeling of intra-vector dependencies.

7. Limitations, Variants, and Future Directions

Vector-wise quantization, as currently deployed, faces several challenges:

• Initialization Sensitivity and Local Minima: EM-style and cluster-based approaches (e.g., KL-minimizing, k-means) may converge to suboptimal partitions, with performance depending on initialization (Yang et al., 2015).
• Scalability: Complexity per iteration for assignment and codebook update scales as $O(NM|\mathcal Y|)$ or worse in large-scale codebook settings, constraining ultra-large models (Yang et al., 2015, Gautam et al., 2023).
• Codebook Collapse: Hard VQ may underutilize codewords; regularization (prior-entropy, soft assignments) and differentiable SCQ/DiVeQ techniques directly address this.
• Hyperparameter Sensitivity: Choice of group dimension $d$, codebook size, and bit allocation per code pose important trade-offs for both accuracy and hardware efficiency.
• Cross-Modal and Multi-Task Generalization: NSNQuant and TurboQuant demonstrate that distributionally aligned VQ can robustly handle out-of-domain or unseen distributions without re-calibration (Son et al., 23 May 2025, Zandieh et al., 28 Apr 2025).

Emerging directions include dynamic discretization per input or layer (DVQ), hybrid scalar/vector methods guided by data uniformity/outliers (RWKVQuant), and fully end-to-end differentiable assignments via convex optimization or reparameterization.

---

References

• "Vector Quantization by Minimizing Kullback-Leibler Divergence" (Yang et al., 2015)
• "SSVQ: Unleashing the Potential of Vector Quantization with Sign-Splitting" (Li et al., 11 Mar 2025)
• "Polar Coordinate Decoupled Vector Quantization" (Yue et al., 5 Jun 2025)
• "Regularized Vector Quantization for Tokenized Image Synthesis" (Zhang et al., 2023)
• "Soft Convex Quantization: Revisiting Vector Quantization with Convex Optimization" (Gautam et al., 2023)
• "vqSGD: Vector Quantized Stochastic Gradient Descent" (Gandikota et al., 2019)
• "TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate" (Zandieh et al., 28 Apr 2025)
• "And the Bit Goes Down: Revisiting the Quantization of Neural Networks" (Stock et al., 2019)
• "Accurate Compression of Text-to-Image Diffusion Models via Vector Quantization" (Egiazarian et al., 2024)
• "VQ4DiT: Efficient Post-Training Vector Quantization for Diffusion Transformers" (Deng et al., 2024)
• "NSNQuant: A Double Normalization Approach for Calibration-Free Low-Bit Vector Quantization of KV Cache" (Son et al., 23 May 2025)
• "RWKVQuant: Quantizing the RWKV Family with Proxy Guided Hybrid of Scalar and Vector Quantization" (Xu et al., 2 May 2025)
• "Intrinsic stationarity for vector quantization: Foundation of dual quantization" (Pagès et al., 2010)
• "DiVeQ: Differentiable Vector Quantization Using the Reparameterization Trick" (Vali et al., 30 Sep 2025)
• "VecQ: Minimal Loss DNN Model Compression With Vectorized Weight Quantization" (Gong et al., 2020)